ROUNDS IN COMMUNICATION COMPLEXITY REVISITED

ROUNDS IN COMMUNICATION COMPLEXITY REVISITED NOAM NISAN AND AVI WIDGERSON Abstract. The k-round two-party communication complexity was studied in the deterministic model by [14] and [4] and in the probabilistic model by [20] and [6]. We present new lower bounds that give (1) randomization is more powerful than determinism in k-round protocols, and (2) an explicit function which exhibits an exponential gap between its k and (k 1)-round randomized complexity. We also study the three party communication model, and exhibit an exponential gap in 3-round protocols that differ in the starting player. Finally, we show new connections of these questions to circuit complexity, that motivate further work in this direction. 1. Introduction. 1.1. The Two-Party Model. Papadimitriou and Sipser [14] initiated the study of how Yao s model 1 [19] of communication complexity is affected by limiting the two players to only k rounds of messages. They considered the following natural problem g k : each of the players A and B is given a list of n pointers (each of log n bits), each pointing to a pointer in the list of the other. Their task is to follow these pointers, starting at some fixed v 0 A, and find the k th pointer. This can easily be done in k rounds and complexity O(k log n): A starts and the players alternately send the value of the next pointer. It is not clear how to use less than n log n bits if only (k 1) rounds are allowed or in fact with k-rounds but player B starts. Indeed, [14] conjectured that the complexity is exponentially higher (for fixed k), namely that there is a strict hierarchy, and proved it for the case k = 2. The general case was resolved by Duris, Galil, and Schnitger [4] who gave an Ω(n/k 2 ) lower bound on the (k 1) round complexity of g k. It is not difficult to see that allowing randomness g k can be solved with high probability in (k 1) rounds using only O((n/k) log n)) communication bits. Another log n factor in the complexity can make this a Las Vegas (errorless) algorithm. This raises the question: what is the relative power of randomness over determinism in k-round protocols? Without limiting the number of rounds Mehlhorn and Schmidt [12] showed This work was partially supported by the Wolfson Research Awards (both authors) and the American-Israeli Binational Science Foundation grant 89-00126 (first author), administered by the Israel Academy of Sciences and Humanities Hebrew University Hebrew University and Princeton University 1 In fact, they and [4] considered the stronger arbitrary partition model, but known simulation results of [4, 6, 9, 10] allow us to use Yao s standard fixed partition model without loss of generality 1

a quadratic gap between Las Vegas and Determinism, and allowing error, the gap can be exponential. We use simple information theoretic and probabilistic arguments to strengthen the lower bound of [4] in two ways. First we improve their (k 1)-round deterministic lower bound on g k to Ω(n) (for k < n/ log n), thus showing that randomness can be cheaper by a factor of k/ log 2 n for k-round protocols. This result also provides the largest gap known for k > log n in the deterministic model - the previous one was obtained in [4] via counting arguments. The fact that the simulation of McGeoch [10] is constructive, gives the same gap in the arbitrary partition model for an explicit function, resolving an open question of [4]. Second, we prove that the probabilistic upper bound above is not very far from optimal we give an Ω(n/k 2 ) lower bound, establishing an exponential gap in the probabilistic setting between k and (k 1)-round protocols for an explicitly given function. The existance of such functions (with somewhat larger gap) was proved by Halstenberg and Reischuk [6], via complicated counting arguments. The only previous exponential gap for an explicit function was shown for k = 2 by Yao [20]. We stress the simplicity of our proof technique, in contrast to that of [6]. We have recently learned the similar techniques were used by Smirnov [16] to prove an Ω(n/(k(log n) k )) lower bound on g k, which is much weaker than our bound, but establish an exponential gap as well. Finally, we use the communication complexity characterization of circuit depth of Karchmer and Wigderson [8] to establish g k as a complete problem for monotone depth-k Boolean circuits. (This result was independently discovered by Yanakakis [18]). Thus a simple deterministic reduction enables to derive the monotone constant-depth hierarchy of Kalwe, Paul, Pippenger and Yannakakis [7] from the constant-round hierarchy of [4]. (The reverse direction was proven in [7]). We speculate that our new probabilistic lower bound may serve to extend the monotone circuit hierarchy result to depth above log n, via probabilistic reductions (as was done by Raz and Wigderson [15]). 1.2. The Multi-Party Model. Chandra, Furst and Lipton [2] devised the multiparty communication complexity model. Here t players P 1, P 2,..., P t are trying to compute a Boolean function g(x 1, x 2,..., x t ), where x i {0, 1} n i. (Until now all work in this model considered equal length inputs, i.e. n i = n for all i). The twist is that 2

every player P i sees all values x j for j i. This model turns out to capture diverse computational models. [2] used it to prove that majority requires superlinear length constant width branching programs. Babai, Nisan and Szegedy [1] gave Ω(n/2 t ) lower bounds for explicit functions g, and used it for Turing machine, branching program and formulae lower bounds, as well as efficient pseudo-random generators for small space. Recently, Håstad and Goldmann [5] used the results in [1] to prove lower bounds on constant-depth threshold circuits. We consider only the 3-player model, and within it allow three rounds of communciation: one per player. We exhibit a function u whose complexity is Ω( n) if P 3 is the first to speak, but O(log n) otherwise. The proof uses properties of universal hash functions developed in [13, 11]. It is interesting that u acts on different size arguments; u : {0, 1} 2n {0, 1} n {0, 1} log n {0, 1} so n 1, n 2 = Θ(n), but n 3 = log n. The following connection to circuit complexity makes such functions important. We show that improving our lower bound to Ω(n) for some explicit function g of this form would give the following size-depth trade-off: the function f : {0, 1} Θ(n) {0, 1} n defined by f(x 1, x 2 ) x3 = g(x 1, x 2, x 3 ) cannot be computed by Boolean circuits of size O(n) and depth O(log n) simultaneously. This result is obtained via Valiant s [17] method of depth-reduction in circuits. 2. The Two-Party Model. The four subsections of this section give the definitions, results, technical lemmas and some proofs, respectively in the two-party communication complexity model. 2.1. Definitions. Let g : X A X B {0, 1} be a function. The players A, B receive respectively inputs x A X A, x B X B. A k-round protocol specifies for each input a sequence of k messages, m 1, m 2,..., m k sent alternately between the players such that at the end both know g(x A, x B ). The cost of a k-round protocol is (where m i is the binary length of m i ), maximized over all inputs (x A, x B ). Denote by C A,k (g) (resp. C B,k (g)) the cost of the best protocol in which player A (resp. B) sends the first message, and C k (g) = min{c A,k (g), C B,k (g)}. Let T : X A X B {0, 1} be the function computed by the two players following a protocol T. We introduce randomization by allowing T to be a random variable distributed over deterministic protocols. The cost is simply the expectation of the associated random variable. We say that randomized protocol makes ɛ-error if 3 k i=1 m i

Pr[T (x A, x B ) g(x A, x B )] ɛ for every input (x A, x B ) X A X B. Denote by C k ɛ (g) the cost of the best k-round ɛ-error protocol for g, and similarly define C A,k ɛ case ɛ = 0 (e.g. C k 0 (g)) denotes Las Vegas (errorless) protocols., C B,k ɛ. The Finally, if we leave T a deterministic protocol, and choose the input uniformly at random, we can define the ɛ-error distributional complexity D k ɛ (g) to be the cost of the best k-round protocol for which Pr[T (x A, x B ) g(x A, x B )] ɛ, under this distribution. The following lemmas are useful. Lemma 2.1. [20] For every g, ɛ > 0 D k 2ɛ(g) 2C k ɛ (g). Lemma 2.2. For all constants 1 3 ɛ > ɛ > 0 C k ɛ (g) = O(Ck ɛ (g)). 2.2. Results. Let V A, V B be two disjoint sets (of vertices) with V A = V B = n and V = V A V B. Let F A = { {f A : V A V B }, F B = {f B : V B V A } and f = (f A, f B ) : fa (v) v V V V defined by f(v) = A. f B (v) v V B For each k 0 define f (k) (v) by f (0) (v) = v, f (k+1) (v) = f(f (k) (v)). Let v 0 V A. The functions we will be interested in computing is g k : F A F B V defined by g k (f A, f B ) = f (k) (v 0 ). Remarks: In the following theorems note that the number of input bits to each player is n log n, and that they hold for every value of k. We also note that one can make g k a Boolean function by taking (say) the parity of the output vertex. All our upper and lower bounds apply to this Boolean function as well. Theorem 2.3. [14] C A,k (g k ) = O(k log n). Theorem 2.4. C B,k (g k ) = Ω(n k log n). Theorem 2.5. C B,k 1/3 (g k) = O((n/k) log n), C B,k 0 (g k ) = O((n/k) log 2 n). Theorem 2.6. C B,k 1/3 (g k) = Ω( n k 2 k log n). Remark: At the end of the proofs of theorems 2 and 4 we explain why the ugly term k log n in these lower bounds can be essentially ignored. In the remainder we show the completeness of g k for monotone depth k circuits. Let g k = g k,n to stress that each player gets n vertices. Definition: For a boolean function h define L d (h) to be the size of the minimal monotone formula of depth d and unbounded fanin that computes h. Define LS d (k, n) to be the maximum of L d (h) over all functions h that can be computed by monotone circuits of unbounded fanin depth k and total size n. Define LF d (k, n) to be the maximum of L d (h) over all functions h that can be computed by a formula of depth k 4

and fanin n at each gate. Theorem 2.7. log LS d (k, n) C d (g k,n ) log LF d (k, n) The left inequality was proven in [7], and allowed them to deduce a lower bound on g k from their circuit lower bound. The right inequality was independently discovered by Yannakakis [18]. It allows to recover the tight hierarchy theorem of [7] from the lower bound on g k. Let h k be the complete function for depth k-circuits, i.e. an alternating and-or tree of depth k and fanin n 1/k at each gate. Corollary 2.8. [7]: Any monotone circuit of depth k 1 for h k requires size 2 Ω(n1/k /k). 2.3. Probability, Measure and Information Theory. Let Ω be a finite set (universe), X Ω. Denote by µ(x) the density of X in Ω, µ(x) = X Ω. Let P : X [0, 1] a probability distribution on X, and x X a random variable distributed according to P. The probability of any event Y X is denoted Pr P [Y ], and the subscript P is usually omitted. For y X, we write Pr[{y}] = P y. Then the entropy H(P ) = H(x) = P y log P y. The information on X (relative to Ω), is I(x) = y X log Ω H(x). If P is the uniform distribution U on X, then H(x) = log X, and I(x) = log µ(x). The following lemmas will be useful to us. Lemma 2.9. Let x = (x 1, x 2,..., x m ) be a random variable (so Ω = Ω 1 Ω 2 Ω m and x i distributed over Ω i ), then I(x) m I(x i ). The next lemma (from [15]) shows that if I(x) is very small, one can get good bounds on the probability of any event under P in terms of its probability under the uniform distribution U. Lemma 2.10. [15] For Y X, let q = Pr U [Y ]. Assume = Pr P [Y ] q q. i=1 4I(x) q 1 10. Then Lemma 2.11. If X = Ω = {0, 1}, I(x) δ 1 4, then P 0 1 2, P 1 1 2 2 δ. 2.4. Proofs. 5

Proof of Theorems 2.3 and 2.5. C A,k (g k ) k log n follows easily, since in round t the right player knows f(v t 1 ) = v t and can send these log n bits to the second player. The idea in beating the determinstic Ω(n) lower bound when the wrong player B starts is as follows: First B chooses a random subset U V B with U = 10n/k, and sends to A {f B (u) : u U}. Now it is A s turn and they start sending each other v 1, v 2,... as above, but lagging one round behind schedule. However, with probability 2/3, one of the v i s will be in U, which allows them to save two rounds, and finish on time. This gives Cɛ B,k (g k ) = O((k+n/k) log n). This algorithm can be made Las-Vegas with an extra factor of O(log n) in the complexity. Proof of Theorems 2.4 and 2.6 Let f = (f A, f B ) F A F B be the input. Let T be a deterministic k-round protocol for g k in which B sends the first message. Note that at any round t 1, if it is B s turn to speak, then v t 1 = f (t 1) (v 0 ) V A, and vice versa. It will be convenient to replace T by a protocol T in which in any round t 1, we replace the message m by the message (m, v t 1 ). By induction on t, this is always possible for the player whose turn it is. In particular, it implies that log n bits are sent per round. Thus if T used C bits, T uses C + k log n bits. We will assume T uses ɛn 2 later), and obtain a contradiction. bits, (ɛ will be chosen Every node z of the protocol tree T can be labeled by the rectangle F z A F z B of inputs arriving at z. By the structure of T, if z is at level t 1 (the root is at level 0), then v 0, v 1,..., v t 1 are determined in F z A F z B. We shall assume the input is chosen uniformly at random from F A F B, so in fact we shall bound from below the distributional comlexity. Thus the probability of arriving at z is µ(f z A F z B), and given that the input arrived at z, it is uniformly distributed in F z A F z B. The main lemma below intuitively shows that if the input arrived at z and the rectangle at z has nice properties, then with high (enough) probability the input will proceed to a child w of z which is equally nice. Nice means that both F z A, F z B are large enough, and that the player not holding v t 1 has very little information on v t = f(v t 1 ). Denote by c z the total number of bits sent by the players before arriving at z. Assume without loss of generality that A speaks at z. Let f z A and f z B be random variables uniformly distributed over F z A and F z B, respectively. Recall that T uses ɛ 2 n bits, and let δ satisfy δ = Max {4 ɛ, 400 ɛ}. Define z to be nice if it satisfies: 6

1. I(fA) z 2c z 2. I(fB) z 2c z 3. I(fB(v z t 1 )) δ Main Lemma: If z is nice, and w a random child of z, then Pr[w not nice] 4 ɛ + 1. n Proof: Assume A sends bits at z. Let a w be the length of the message which leads to the child w. Thus c w = c z + a w for all children w of z. We will now give upper bounds separately on the probability of each of the three properties defining nice being false at a random child w. Claim 1. Pr[I(fB w ) > 2c w ] = 0. Proof: B sent nothing, so w FB w = FB z and I(f w B ) = I(f z B) 2c z < 2c w. Claim 2. Pr[I(f w A ) > 2c w ] 1 n. Proof: A child w is chosen with probability µ(f w A ) / µ(f z A). Also note that the for every w a w log n (by our assumption on the protocol) and they satisfy Kraft s inequality w 2 aw 1 (where the sum is over all children of z. Thus we have Pr[I(f w A ) > 2c w ] Pr[µ(F w A ) < 2 2cw ] Pr[µ(FA w ) / µ(fa) z < 2 2aw ] 1 2 aw 1 n w n. Claim 3. Pr[I(f w A (v t )) > δ] 4 ɛ. Proof: We may assume now that I(fA w ) 2c w ɛn. The random variable fa w is a vector of random variables fa w (v) for all v V A. Thus by Lemma 2.9, I(fA w (v)) I(fA w ) ɛn. So if v t was chosen uniformly from V A, Pr U [I(fA w (v t )) > δ] ɛ by Markov s δ inequality. But v t = fb(v z t 1 ), so v t is distributed with I(v t ) = I(fB(v z t 1 )) δ as we assumed z was nice. By Lemma 2.10 (and our choice of δ), Pr[I(f w A (v t )) > δ] ɛ δ ( 1 + ) 4δ 4 ɛ. ɛ/δ Now we can conclude the proofs of Theorems 2.4 and 2.6 from the main lemma. Consider any nice leaf l of the protocol tree T, labeled by an answer (0 or 1). Say A spoke on the last round k. Then I(v k ) = I(f l B(v k 1 )) δ. So by Lemma 2.11, even 7 v V

if the algorithm gives one bit (say parity) of the answer, it is correct with probability 1 2 + 2 δ. Conclusion of Theorem 2.4 Take ɛ = 10 4. The root of T is nice, so by the main lemma and induction we have a positive probability ( 2 k ) of reaching a nice leaf, contradicting the fact that the protocol never errs. This proves only C B,k (g k ) = Ω(n k log n), since we augmented an arbitrary T to a nice protocol T. Getting rid of the k log n term, achieving the lower bound C B,k (g k ) = Ω(n) (which is stronger when k n ) requires a more delicate argument that we sketch below. The log n idea is to follow the same steps of the proof with the following changes. (1) We stay with the original protocol T, as we cannot afford the players sending log n bits per round as in the nice protocol T. (2) We still fix the vertex v t 1 by the player sending the message at round t, but avoid paying log n bits for this information by removing this vertex from our universe. Thus the information I is measured relative to a smaller set of pointers at every round. (3) We prove a weaker main lemma, which is clearly sufficient in the deterministic case, namely that every nice node z has at least one nice child w. The details are left to the interested reader. Conclusion of Theorem 2.6. Pick ɛ = 10 4 k 2. Thus the probability of not reaching a nice leaf is k 1 = 1, and the probability that the protocol answers 25k 25 ) < 0.95. Thus we get from Lemmas 2.1 and 2.2 that correctly is less than 1 25 + ( 1 2 + 2 5 k D B,k 1/20 (g k) = Ω( n k 2 k log n). This bound is Ω( n k 2 ) for all k < ( n log n )1/3. For larger k one can use the trivial lower bound k which applies to every k-round protocol. Note that when k n 1/3 this trivial bound is larger than n/k 2. Proof of Theorem 2.7: As mentioned above, the left inequality was proven in [7], so we prove only the right inequality. The proof is based on the Karchmer-Wigderson characterization of circuit depth in terms of communication complexity, which can be stated as follows. For every monotone function h on n variables with minterms M in(h) and maxterms Max(h) define a communication search problem R m h Min(h) Max(h) [n] in which player A gets a minterm S Min(h), player B gets a maxterm T Max(h), and their task is to find an element in S T. protocols for R m h Then monotone formulae for h and are in 1-1 correspondence via the simple syntactic identification of gates with player A s moves and gates with player B s moves. In particular, depth corresponds to the number of rounds, and logarithm of the size to the communication 8

complexity. In view of the above, all we need to give now is a reduction from computing g k,n to the computation of Rh m for some function h which has a depth k formula of fanin n at each gate. Once this is done the players can solve Rh m and hence g k,n in d rounds and loglf d (k, n) communication by simulating the guaranteed depth d circuit for h. Let h be defined by a formula that is a complete n-ary tree of depth k, alternating levels of and gates (say with at the root), and distinct n k variables at the leaves. The players agree on a fixed labeling of the nodes of this tree in which the root is labeled v 0, the children of every gates labeled by V B, and children of every gate labeled by V A. Let f A and f B be the inputs to players A, B respectively. Player A constructs sets S i of nodes from the ith level inductively as follows. S 0 contains the root. If level i contains gates, then for every gate in S i labeled v he adds to S i+1 the unique child of this gate labeled f A (v). If level i contains gates, then for every gate in S i he adds all its children to S i+1. In a similar way (exchanging the roles of gates) player B constructs his sets T i. It is easy to verify that S k is a minterm of h, T k is a maxterm of h, and that they intersect at a unique leaf, whose label is f (k) (v 0 ). This completes the reduction, and hence the proof. 3. The Three-Party Model. Let g : {0, 1} n 1 {0, 1} n 2 {0, 1} n 3 {0, 1} be a function. Players P 1, P 2, P 3 are given (x 2, x 3 ), (x 1, x 3 ), (x 1, x 2 ) respectively with x i {0, 1} n i and compute g from this information by exchanging messages according to a predetermined protocol. We consider only 3-round protocols in which each player speakes once. Let M i (g) denote the communication complexity when player P i speaks first (and then the other two in arbitrary order), and M s (g) the complexity when they all speak simultaneously (an oblivious protocol). Clearly, for all i {1, 2, 3} M i (g) M s (g). Let u : {0, 1} 2n {0, 1} n {0, 1} log n {0, 1} be the following function. Interpret the first string x 1 as a 2-universal hash function ([3]) h, mapping {0, 1} n to itself, the second string x 2 as an argument y to h, and the third x 3 as an index j [n]. Then u(h, y, j) = h(y) j. The next two theorems exhibit an exponential gap between 3-round protocols that differ in the order in which players speak. Theorem 3.1. M 1 (u) = M 2 (u) = O(log n) Theorem 3.2. M 3 (u) = Ω( n). 9

Proof of Theorem 3.2 We first recall a fundamental lemma from [11] regarding the distribution of hash values given little information on the hash function and the argument. Lemma 3.3. [11] Let H = {h : I O} be a collection of universal hash functions from domain I into range O. Let A I, B O, C H and p = B / O. Then Pr[h(x) B x A, h C] p p H /( A C ) Restrict the value of j to be j [ n]. Thus we consider h : {0, 1} n {0, 1} n which is still a universal hash function. Assume M 3 (u) n/5. This means that there is a new protocol (independent of j) to compute z = h(y) in which P 3 sends n/5 bits, and then players P 1 and P 2 can compute each bit of z separately, using altogether n/5 bits. Let (h, y) be chosen uniformly at random. Simple averaging shows that there are messages m 1, m 2, m 3 of P 1, P 2, P 3, respectively, which under this distribution satisfy Pr[m 1 ] 2 n/4, Pr[m 2 ] 2 n/4, and Pr[m 3 ] 2 n/4. As m 1 coresponds to a subset C of all hash functions (inputs to P 1 ), and m 2 corresponds to a subset A of all inputs to P 2, we can use Lemma 3.3 with H / C 2 n/4, 1/ A 2 3n/4, B = {z} and p = 2 n to obtain Pr[h(y) = z m 1, m 2, m 3 ] 2 n/4 Pr[h(y) = z m 1, m 2 ] 2 n/4 2 n/2 = 2 n/4 < 1. Let f : {0, 1} m {0, 1} n be an arbitrary function, and for any m < m define g f : {0, 1} m {0, 1} m m {0, 1} log n {0, 1} by g f (x 1, x 2, x 3 ) = f(x 1 x 2 ) x3, where denotes concatenation. The next theorem gives the relationship of size-depth trade-offs in circuits to 3-round oblivious protocols. Theorem 3.4. If f above can be computed by a circuit of fan-in 2, size O(n) and depth O(log n), then M s (g f ) = O(n/ log log n). Proof of Theorem 3.4 Let f : {0, 1} m {0, 1} n be computed by a circuit C of size O(n) and depth O(log n). By a result of Valiant [17], there are s = O(n/ log log n) wires in C, e 1,e 2,...,e s with the following property. For every input x {0, 1} m, and every j [n], f(x) j is determined by the values e 1 (x),..., e s (x) on these wires, together with the values of x i, i S j with S j n. To compute g f, note that P 3 has access to x = (x 1 x 2 ) 10

(which is the input to f) and therefore can compute the values on the wires. P 2 and P 3, now knowing j = x 3, exchange the necessary bits in S j to complete the computation of f(x) j. Acknowledgements: We are greatful to the referees for careful reading and many valuable suggestions. 11

REFERENCES [1] L. Babai, N. Nisan, M. Szegedy: Multiparty protocols and logspace-hard pseudorandom sequences Proc. of the 21st STOC (1989) 1-11. [2] A. Chandra, M. Furst, R. Lipton: Multi-parti Protocols Proc. of the 15th STOC (1983) 94-99. [3] L. Carter, M. Wegman: Universal hash functions Journal of Comp. and Sys. Sci. 18 (1979) 143-154. [4] P. Duris, Z. Galil, G. Schnitger: Lower Bounds of Communication Complexity Proc. of the 16th STOC, (1984) 81 91 [5] J. Håstad, M. Goldmann: On the power of small depth threshold circuits Proc. of the 31st FOCS, (1990) 610-618. [6] B. Halstenberg, R. Reischuk: On Different Modes of Communication Proc. of the 20th STOC (1988) 162-172. [7] M. Klawe, W.J. Paul, N. Pippenger, M. Yannakakis: On Monotone Formulae with Restricted Depth Proc. of the 16th STOC, (1984) 480 487 [8] M. Karchmer, A. Wigderson: Monotone Circuits for Connectivity Require Super-Logarithmic Depth Proc. of the 20th STOC, (1988), 539 550 [9] T. Lam, L. Ruzzo: Results on Communication Complexity Classes Proc. of the 4th Structures in Complexity Theory conference, (1989) 148-157. [10] L.A. McGeoch: A Strong Seperation Between k and k 1 Round Communication Complexity for a Constructive Language CMU Technical Report CMU-CS-86-157 (1986) [11] Y. Mansour, N. Nisan, P. Tiwary: The computational complexity of universal hashing Proc. of the 22nd STOC (1990). [12] K. Mehlhorn, E. Schmidt: Las Vegas is better than Determinism in VLSI and Distributed Computing Proc. of the 14th STOC (1982), 330-337. [13] N. Nisan: Pseudorandom Generators for Space Bounded Computation Proc. of the 22nd STOC, (1990) 204-212. [14] P.H. Papadimitriou, M. Sipser: Communication Complexity Proc. of the 14th STOC, (1982) 330 337 [15] R. Raz, A. Wigderson: Probabilistic Communication Complexity of Boolean Relations Proc. of the 30th FOCS, (1989) 562 567 [16] D. V. Smirnov: Shannon s information methods for lower bounds for probabilistic communication complexity, Manuscript (in Russian), (1989) [17] L. Valiant: Graph theoretic arguments in low-level complexity, Technical Report CS 13-77, University of Edinburgh (1977). [18] M. Yannakakis: Private communication. [19] A.C.-C. Yao: Some Complexity Questions Related to Distributive Computing, Proc. of the 11th STOC, (1979) 209 213 [20] A.C.-C. Yao: Lower Bounds by Probabilistic Arguments, Proc. of the 24th FOCS, (1983) 420 428 12